High-voltage switch cabinet partial discharge fault identification method

Abstract

The invention discloses a high-voltage switch cabinet partial discharge fault identification method and a storage medium. The method comprises the following steps: synchronously acquiring acoustic, electric, magnetic and thermal signals through a multi-physical field sensor array; performing preprocessing and feature extraction on the signals; using a cascade deep learning network to fuse multi-modal features to identify fault types and levels; realizing accurate positioning of the discharge source in the digital twinborn model by adopting a dynamic simulation coherent positioning method; and risk assessment and early warning are carried out by integrating diagnosis and positioning results. According to the method, deep learning and physical mechanism simulation are deeply fused through a data and model dual-drive normal form, so that the defects of shallow fusion level, low positioning precision, lack of interpretability and the like in the prior art are overcome, and intelligent, precise and prospective identification and early warning of the partial discharge fault of the high-voltage switch cabinet are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring and fault diagnosis, and particularly relates to a high-voltage switch cabinet partial discharge fault identification method. BACKGROUND

[0002] The high-voltage switch cabinet is a crucial power distribution equipment in the power grid, and the degradation of its internal insulation state is usually preceded by partial discharge. Timely and accurate detection and positioning of partial discharge is of great significance for preventing equipment failure and ensuring safe operation of the power grid.

[0003] Existing partial discharge detection methods mostly rely on multi-sensor technology, such as ultrasonic method, ultra-high frequency method and pulse current method. In terms of information processing and fault identification, the traditional approach is to use algorithms such as BP neural network to fuse and identify features, and to use the time difference of arrival (TDOA) method for positioning. However, these methods have obvious limitations: first, shallow fusion models such as BP neural network have difficulty in fully mining deep temporal and spatial correlation features in multi-modal data (such as images and sequences), resulting in limited recognition ability for complex discharge patterns. Second, the TDOA positioning method is based on the ideal model of "straight-line propagation of sound waves", which cannot effectively handle the problems of sound wave reflection, refraction and speed variation caused by the complex structure inside the switch cabinet, resulting in large positioning errors. In addition, existing methods are mostly single paradigms of "data-driven" or "model-driven", lacking means to deeply integrate physical mechanisms and data intelligence, resulting in poor interpretability and weak generalization ability of the system.

[0004] Therefore, there is an urgent need in the art for an innovative method that can deeply integrate physical mechanisms and data intelligence, achieve accurate identification and positioning, and have forward-looking early warning capability. SUMMARY

[0005] The purpose of the present application is to provide a high-voltage switch cabinet partial discharge fault identification method to solve the problems of shallow multi-modal information fusion level, positioning accuracy of discharge source limited by ideal model, and lack of risk forward-looking assessment in the prior art.

[0006] To achieve the above purpose, the present application adopts the following technical solutions: A high-voltage switch cabinet partial discharge fault identification method, characterized in that it comprises the following steps: S100, synchronously collecting ultrasonic signals, ultra-high frequency electromagnetic wave signals, pulse current signals and temperature signals generated by partial discharge through a multi-physical field sensor array deployed inside the high-voltage switch cabinet; S200: preprocessing and feature extraction of the collected multi-modal signals; S300: based on the cascaded deep learning network, the extracted multi-modal features are analyzed, the fault type of partial discharge is identified, and the severity level is evaluated; the cascaded deep learning network comprises a convolutional neural network for processing time-frequency image features, a gated recurrent unit network for processing time series signal features, and a multi-head self-attention mechanism for feature weighted fusion; S400: based on the dynamic simulation coherent positioning method, in the digital twin model of the high-voltage switch cabinet, the position where the simulation signal and the measured signal have the maximum coherence is searched by an optimization algorithm, and the three-dimensional accurate positioning of the partial discharge source is realized; S500: the fault diagnosis and early warning report is generated by comprehensively analyzing the fault type, severity level and positioning information.

[0007] The application further provides that the multi-modal signals are preprocessed and features are extracted, specifically including: S200-1: the adaptive noise complete ensemble empirical mode decomposition method is used to denoise the original ultrasonic signal and pulse current signal; S200-2: the denoised ultrasonic signal and ultrahigh frequency signal are subjected to continuous wavelet transform to generate a two-dimensional time-frequency graph as an image modal feature; S200-3: the denoised pulse current signal is cut to obtain time series data of multiple power frequency periods as sequence modal features.

[0008] The application further provides that in step S300, the fusion analysis based on the cascaded deep learning network specifically includes: S300-1: the time-frequency graph generated by the ultrasonic wave and the ultrahigh frequency signal is respectively input into two weight-shared convolutional neural network branches to extract spatial-spectral features; S300-2: the time series data of the pulse current is input into the gated recurrent unit network to extract time dynamic features; S300-3: the feature vectors output by the two CNN branches and the feature vectors output by the GRU network are spliced and sent to the multi-head self-attention mechanism layer for weighted fusion to adaptively emphasize key features; S300-4: based on the fused features, the fault type, severity level and diagnostic confidence are output by the classifier.

[0009] The application further provides that in step S400, the following steps are further included: S400-1: a digital twin model of the high-voltage switch cabinet is constructed, which integrates three-dimensional geometry, material physical properties and temperature field distribution; S400-2: a virtual discharge source candidate position is set in the digital twin model; S400-3: based on the physical attribute of the digital twin model, simulate and calculate the simulation waveform of the ultrasonic wave propagating from the candidate position to each sensor; S400-4: calculate the average coherence in the frequency domain of the measured ultrasonic wave signal and the simulation waveform on each sensor channel; S400-5: taking the maximum average coherence as the objective function, iteratively updating the candidate position of the virtual discharge source by using the particle swarm optimization algorithm; S400-6: determining the optimal candidate position at the iteration termination as the actual position of the partial discharge source.

[0010] The application further provides that between step S300 and step S400, step S600 is further included, that is, updating the cascade deep learning network based on a federated learning framework; wherein, multiple edge nodes train local models on local data and upload model weights to the cloud, and the cloud server generates a global model after securely aggregating the weights and distributes the global model to each edge node.

[0011] The application further provides that in step S500, a risk assessment step is further included: S500-1: receiving the fault type, severity level and positioning information; S500-2: combining device historical working condition data, real-time load and environmental factors, and calculating short-term and long-term risk probabilities of fault evolution into accidents by a Bayesian risk network; S500-3: automatically generating a differentiated operation and maintenance strategy based on the risk probability.

[0012] The application further provides that in step S400-1, the digital twin model receives real-time temperature data and updates the temperature field distribution inside the model, and further dynamically adjusts the ultrasonic wave propagation speed in acoustic simulation.

[0013] The application further provides that in step S200-1, the adaptive noise complete set empirical mode decomposition method is used to decompose the signal into multiple intrinsic mode functions, and the IMF components related to the partial discharge characteristics are screened and reconstructed to realize signal denoising.

[0014] The application further provides that in step S100, the collection process of the multi-physical field sensor array is controlled by a precise clock synchronization unit, ensuring that the synchronization accuracy of all signals is in the microsecond level.

[0015] The application further provides a computer readable storage medium having a computer program stored thereon, when the program is executed by a processor, the steps of the method are realized.

[0016] The beneficial effects of the present application: compared with the prior art, the technical scheme provided by the present application brings the following significant beneficial effects: Deep fusion, accurate diagnosis: through the cascade network of CNN+GRU+Attention, deep and adaptive fusion of multi-modal data such as images and sequences is realized, and the recognition accuracy and robustness of complex discharge modes are significantly improved.

[0017] Mechanism embedding, positioning innovation: dynamic simulation coherent positioning method deeply embeds physical mechanism in the positioning process, fundamentally overcomes the model defects of traditional TDOA method, realizes the leap of positioning accuracy from "decimeter level" to "centimeter level", and has physical interpretability.

[0018] Strong anti-interference capability: the application of advanced signal processing technologies such as CEEMDAN noise reduction and frequency domain coherence matching makes the method still maintain high reliability in strong noise industrial environment.

[0019] Intelligent early warning, decision foresight: the introduction of Bayesian risk network realizes the leap from fault diagnosis to risk prediction, provides quantitative basis for predictive maintenance, and greatly improves the pertinence and economy of operation and maintenance activities.

[0020] Continuous evolution, protect privacy: the introduction of federated learning mechanism enables the method to use global data wisdom for continuous optimization, while strictly protecting the privacy and security of original data, breaking the data island.

[0021] Excellent generalization ability: weight sharing CNN, federated learning and other designs ensure that the method can maintain excellent performance when facing new environments and new devices, and has strong generalization application potential. BRIEF DESCRIPTION OF DRAWINGS

[0022] The drawings described herein are used to provide further understanding of the present application, constitute a part of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitation on the present application. In the drawings: Figure 1 The flowchart of the embodiment of the present application is shown. DETAILED DESCRIPTION

[0023] The embodiments of the present application will be described in detail below with the help of drawings and examples, so that the realization process of applying technical means to solve technical problems and achieving technical effects of the present application can be fully understood and implemented.

[0024] As shown in Figure 1 The present application provides a high-voltage switch cabinet partial discharge fault recognition method, which comprises the following steps: S100, synchronously collecting ultrasonic wave signals, ultra-high frequency electromagnetic wave signals, pulse current signals and temperature signals generated by partial discharge through a multi-physical field sensor array deployed inside the high-voltage switch cabinet; S200, preprocessing and extracting features from the collected multi-modal signals; S300, performing fusion analysis on the extracted multi-modal features based on a cascaded deep learning network, identifying the fault type of the partial discharge and evaluating its severity level; the cascaded deep learning network includes a convolutional neural network for processing time-frequency image features, a gated recurrent unit network for processing time series signal features, and a multi-head self-attention mechanism for feature weighted fusion; S400, based on the dynamic simulation coherence positioning method, in the digital twin model of the high-voltage switch cabinet, the position where the simulation signal and the measured signal have the maximum coherence is searched by an optimization algorithm, and the three-dimensional accurate positioning of the partial discharge source is realized; S500, generating a fault diagnosis and warning report by integrating the fault type, severity level and positioning information.

[0025] Wherein the preprocessing and feature extraction of the multi-modal signals specifically include the following steps: S200-1, using an adaptive noise complete ensemble empirical mode decomposition method to denoise the original ultrasonic wave signals and pulse current signals; S200-2, performing continuous wavelet transform on the denoised ultrasonic wave signals and ultra-high frequency signals to generate two-dimensional time-frequency images as image modal features; S200-3, for the denoised pulse current signals, a plurality of time series data of power frequency cycles are intercepted as sequence modal features.

[0026] It is worth mentioning that in step S300, the fusion analysis based on the cascaded deep learning network specifically includes: S300-1, inputting the time-frequency images generated by the ultrasonic wave and ultra-high frequency signals into two weight-shared convolutional neural network branches respectively to extract spatial-spectral features; S300-2, inputting the time series data of the pulse current into the gated recurrent unit network to extract time dynamic features; S300-3, concatenating the feature vectors output by the two CNN branches with the feature vectors output by the GRU network, and sending them to the multi-head self-attention mechanism layer for weighted fusion to adaptively emphasize key features; S300-4, based on the fused features, outputting the fault type, severity level and diagnostic confidence through the classifier.

[0027] Wherein step S400 includes the following steps: S400-1: Construct a digital twin model of the high-voltage switch cabinet, which integrates three-dimensional geometry, material physical properties, and temperature field distribution; S400-2: Set a virtual discharge source candidate position in the digital twin model; S400-3: Based on the physical properties of the digital twin model, simulate the simulation waveform of the ultrasonic wave propagating from the candidate position to each sensor; S400-4: Calculate the average coherence in the frequency domain of the measured ultrasonic wave signal and the simulation waveform on each sensor channel; S400-5: Take the maximum average coherence as the objective function, and use the particle swarm optimization algorithm to iteratively update the virtual discharge source candidate position; S400-6: Determine the optimal candidate position at the end of iteration as the actual position of the partial discharge source.

[0028] As preferred, between step S300 and step S400 of the present application, step S600 is further included, which updates the cascaded deep learning network based on a federated learning framework; wherein multiple edge nodes train local models on local data and upload model weights to the cloud, and the cloud server generates a global model after securely aggregating the weights and distributes it to each edge node.

[0029] In addition, in step S500, a risk assessment step is further included: S500-1: Receive the fault type, severity level, and positioning information; S500-2: Combine device historical operating condition data, real-time load, and environmental factors to calculate short-term and long-term risk probabilities of fault evolution into accidents through a Bayesian risk network; S500-3: Based on the risk probability, automatically generate differentiated operation and maintenance strategies.

[0030] In step S400-1, the digital twin model receives real-time temperature data and updates the temperature field distribution inside the model, and then dynamically adjusts the ultrasonic wave propagation speed in the acoustic simulation.

[0031] In step S200-1, the adaptive noise complete ensemble empirical mode decomposition method is used to decompose the signal into multiple intrinsic mode functions, and the IMF components related to the partial discharge feature are selected for reconstruction to achieve signal denoising.

[0032] In step S100, the acquisition process of the multi-physical field sensor array is controlled by a precise clock synchronization unit to ensure the synchronization accuracy of all signals within microseconds.

[0033] Now the content of the present application will be further explained: First, the system is deployed and parameterized, and the sensor array is deployed inside the target switch cabinet: 4 ultrasonic sensors (frequency range 40 kHz ± 5 kHz); 2 ultrahigh frequency sensors (frequency range 300 MHz-1.5 GHz); 1 high-frequency current transformer (bandwidth 100 MHz); 3 digital temperature sensors; Edge computing gateway configuration: CPU 4-core 2.5 GHz, memory 8 GB, connected to the cloud platform through industrial Ethernet.

[0034] Then signal acquisition and preprocessing are performed: When the system detects abnormal signals, the diagnosis process is triggered. The original ultrasonic signal is denoised by CEEMDAN, and the process is as follows: the original signal , and then a set of noisy signals is generated, whose formula is as follows: , where is white noise, is the noise intensity,

[0035] Then each is decomposed by EMD to obtain the first modal component , and then the first residual is calculated, and is decomposed by EMD to obtain the second modal component: , where represents the kth modal obtained by EMD decomposition of the signal. Repeat the above steps until the residual is a monotonic function or the number of extreme points is less than 2. In general, the denoising process is as follows:

[0036] In the above formula, is the residual signal after the k-1th decomposition; is the Gaussian white noise added in the ith time; is the adaptive noise coefficient, generally taken as 0.2; is the EMD decomposition operator; N=100 is the set number; after three layers of decomposition, the signal-to-noise ratio is improved from the original 8.2 dB to 26.5 dB.

[0037] Then the denoised ultrasonic and ultrahigh frequency signals are subjected to continuous wavelet transform to generate 256x256 time-frequency images as the input of the CNN network. The continuous wavelet transform convolves the signal x(t) with the wavelet function ψ(t):

[0038]

[0039] where a is a scale parameter, where , is the sampling frequency of the ultrasonic signal, is the minimum effective frequency of the ultrasonic sensor; is the center frequency, generally taken as 6.0; b is a translation parameter, corresponding to the propagation time of the discharge pulse in the switch cabinet, , is the time resolution, corresponding to a sampling rate of 1 MHz. ψ* is the complex conjugate of the wavelet function; The preprocessed multi-modal data is then input into a cascaded deep learning network, where the CNN branch adopts a 3-layer convolution structure with filter numbers of 32, 64, and 128 respectively, and extracts common features through a weight sharing mechanism.

[0040] The three layers of convolution are as follows: low-layer convolution kernels (1st layer) for detecting edge features: the start and end boundaries of the discharge pulse; detecting texture features: the energy distribution patterns in the time-frequency graph; middle-layer convolution kernels (2nd layer) for detecting combined features: the combined patterns of multiple basic features, detecting directional features: the directionality of energy propagation in the time-frequency graph; high-layer convolution kernels (3rd layer) for detecting semantic features: the feature patterns of specific discharge types, detecting abstract features: discharge laws across time scales; Feature extraction process: For the ultrasonic input, the calculation formula is: ; For the ultrahigh frequency input, the calculation formula is: ; Correspondence between convolution kernel functions and partial discharge features Discharge type feature mapping: internal discharge: periodic pulse groups appear in the time-frequency graph; surface discharge: characteristics of energy propagation along the surface; corona discharge: randomly distributed single pulses.

[0041] where the convolution layer formula is: , m, n are the coordinates inside the convolution kernel, is the activation value of the output feature map of the l layer at position (i, j), is the value of the input feature map of the l -1 layer at position (i+m, j+n), is the bias term; The weight sharing constraint is as follows: ; represents the Layer ultrasound wave branch convolution weight. Representing the first Layer ultrahigh frequency branch convolution weight. Representing the first Layer shared convolution weight.

[0042] GRU branch: hidden layer dimension 128, process power frequency cycle sequence data of pulse current, where GRU has two gates, reset gate And update gate . The formula is as follows: ; ; Where the candidate hidden state is: ; Hidden state: Where, σ is the sigmoid function, and * represents element-wise multiplication.

[0043] Attention fusion: 8 heads of self-attention mechanism calculate feature weights, and output fusion results.

[0044] Then the features extracted by CNN and GRU are fused, and the attention weight is calculated as: The scoring function is: Where, , , Respectively from the feature vectors of ultrasonic waves, ultrahigh frequency and pulse current.

[0045] Subsequently, the discharge source position is accurately located by dynamic simulation coherence, and the discharge source is accurately located at a centimeter level based on the digital twin model. Specifically, initialization: the system obtains the measured signals from the four ultrasonic sensors , , , , a candidate position P of the discharge source is proposed by the SPO optimization algorithm, the digital twin model performs acoustic simulation at the position P, and four corresponding simulation signals , , , , coherence comparison: for each sensor i, the coherence coefficient of the measured signal and the simulation signal in the frequency domain is calculated:

[0046] Where, The self-power spectrum of the measured signal represents the energy frequency distribution of the measured signal; for the auto-power spectrum of the simulation signal, representing the energy frequency distribution of the simulation signal; for the cross-power spectrum of the measured signal and the simulation signal, representing the correlation degree of the two signals at different frequencies.

[0047] Calculate the average coherence C(P) of all sensors. The optimization algorithm adjusts the candidate position P according to this "matching score", and repeats the above steps until the optimal position P that maximizes C(P) is found optimal .

[0048] At the same time, considering the influence of the internal structure of the switch cabinet on the propagation of ultrasonic waves, the positioning accuracy is improved through sound wave propagation path correction, and the steps are as follows: Step 1: Digital twin model construction Based on the accurate CAD drawings of high-voltage switch cabinets, a three-dimensional digital twin model is constructed. This model not only contains geometric structures, but also integrates a material attribute database, giving each material in the cabinet (such as air, epoxy resin, copper, SF6 gas, etc.) accurate physical parameters. At the same time, the model receives real-time data from temperature sensors deployed at multiple key points in the cabinet, generating a dynamic three-dimensional temperature field in the entire cabinet space through spatial interpolation algorithms.

[0049] Step 2: Sound wave propagation path analysis based on ray tracing In this step, the system performs a detailed physical simulation of the propagation process of sound waves from the candidate discharge source position to each sensor. Specifically: First, path generation and identification: using the ray tracing algorithm, sound rays are emitted from the candidate discharge source position to each sensor position in the digital twin model. This algorithm intelligently identifies and calculates all major propagation paths, including direct straight-line paths that penetrate directly, and indirect paths that are reflected or diffracted one or more times by various obstacles (such as insulators, metal partitions) in the cabinet before reaching the sensor.

[0050] Second, propagation path segmentation: for each identified propagation path, the system will accurately segment it into multiple small segments according to the different media (materials) it passes through and the different temperature regions it passes through. The fundamental principle of segmentation is to ensure that within each small segment, the medium for sound wave propagation is uniform, and temperature changes are negligible. For example, a path that starts from the discharge source, first penetrates through air, then penetrates through an epoxy resin insulating partition, and finally propagates to the sensor in the air, will be divided into three independent path segments accordingly.

[0051] Finally, segmented parameter calculation: for each path segment obtained by segmentation, the system will perform the following calculations: Geometric length calculation: based on the three-dimensional coordinates of the digital twin model, the actual spatial length of the path segment is calculated.

[0052] Sound speed determination: According to the material type marked on the path segment and the real-time temperature data of the region, the corresponding sound speed model is called (for example, the temperature-dependent sound speed formula is used for the air segment, and the inherent sound speed value is used for the solid material segment), to determine the propagation speed of the sound wave in the segment path.

[0053] Propagation time calculation: According to the formula "time=length / velocity", the time required for the sound wave to pass through the specific path segment is calculated.

[0054] Step 3: Propagation time calculation: The propagation times of all segments contained in a complete propagation path obtained in step 2 are added up, and the additional time delay caused by diffraction, reflection and other physical phenomena is added, so that the total simulation time of the sound wave from the candidate discharge source position to the specified sensor can be obtained. This time is the basis for subsequent simulation waveform generation and coherence calculation.

[0055] Step 4: Comparison with measured time: The theoretical total propagation time calculated in step 3 is compared with the arrival time of the ultrasonic wave signal received by the actual sensor to calculate the time error. This error value is used as one of the key indicators to evaluate the accuracy of the current candidate discharge source position, and is fed back to the optimization algorithm to guide the next position guess.

[0056] The present application has the following advantages: Deep fusion, accurate diagnosis: Through the cascade network of CNN+GRU+Attention, deep and adaptive fusion of multi-modal data such as images and sequences is realized, which significantly improves the recognition accuracy and robustness of complex discharge patterns.

[0057] Mechanism embedding, positioning innovation: The dynamic simulation coherence positioning method deeply embeds the physical mechanism in the positioning process, fundamentally overcoming the model defects of traditional TDOA methods, achieving a leap in positioning accuracy from "decimeter level" to "centimeter level", and having physical interpretability.

[0058] Strong anti-interference capability: The application of advanced signal processing technologies such as CEEMDAN noise reduction and frequency domain coherence matching makes the method still maintain high reliability in strong noise industrial environment.

[0059] Intelligent early warning, decision foresight: The introduction of Bayesian risk network realizes the leap from fault diagnosis to risk prediction, providing a quantitative basis for predictive maintenance, and greatly improving the pertinence and economy of operation and maintenance activities.

[0060] Continuous evolution, protect privacy: the introduction of the mechanism of federated learning enables the method to continuously optimize the wisdom of global data, while strictly protecting the privacy and security of the original data, breaking the data silos.

[0061] Excellent generalization ability: the design of weight-sharing CNN and federated learning ensures that the method can maintain excellent performance when facing new environments and new devices, and has strong generalization potential.

[0062] As some terms are used in the description and claims to refer to certain components. Those skilled in the art should understand that hardware manufacturers may use different names to refer to the same component. The description and claims of this specification do not take the difference in name as the basis for distinguishing components, but take the difference in function as the criterion for distinguishing. As mentioned throughout the description and claims, those skilled in the art can solve the technical problem within a certain error range and basically achieve the technical effect.

[0063] The above description shows and describes several preferred embodiments of the present application, but as mentioned above, the present application is not limited to the form disclosed herein, should not be considered as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the invention described herein by the above teachings or related art or knowledge. The changes and modifications made by those skilled in the art without departing from the spirit and scope of the present application shall be within the scope of protection of the appended claims of the present application.

Claims

1. A partial discharge fault identification method for a high voltage switchgear, characterized in that, The method comprises the following steps: S100, synchronously collecting ultrasonic wave signals, ultra-high frequency electromagnetic wave signals, pulse current signals and temperature signals generated by partial discharge through a multi-physical field sensor array deployed inside a high-voltage switch cabinet; S200, preprocessing and extracting features of the collected multi-modal signals; S300, performing fusion analysis on the extracted multi-modal features based on a cascaded deep learning network, identifying the fault type of the partial discharge and evaluating its severity level; the cascaded deep learning network comprises a convolutional neural network for processing time-frequency image features, a gated recurrent unit network for processing time series signal features, and a multi-head self-attention mechanism for feature weighted fusion; S400, based on a dynamic simulation coherence positioning method, in the digital twin model of the high-voltage switch cabinet, iteratively searching for a position that maximizes the coherence between the simulation signals and the measured signals through an optimization algorithm, to realize three-dimensional accurate positioning of the partial discharge source; S500, generating a fault diagnosis and warning report by comprehensively considering the fault type, severity level and positioning information.

2. The partial discharge fault identification method for a high-voltage switch cabinet according to claim 1, characterized in that, In step S200, the multi-modal signals are preprocessed and features are extracted, specifically including: S200-1, using an adaptive noise complete ensemble empirical mode decomposition method to denoise the original ultrasonic wave signals and pulse current signals; S200-2, performing continuous wavelet transform on the denoised ultrasonic wave signals and ultra-high frequency signals to generate two-dimensional time-frequency images as image modal features; S200-3, for the denoised pulse current signals, time series data of multiple power frequency periods are intercepted as sequence modal features.

3. The partial discharge fault identification method for a high-voltage switchgear according to claim 1, characterized in that, In step S300, the fusion analysis based on the cascaded deep learning network specifically includes: S300-1, inputting the time-frequency images generated by the ultrasonic wave and ultra-high frequency signals into two weight-shared convolutional neural network branches respectively to extract spatial-spectral features; S300-2, inputting the time series data of the pulse current into a gated recurrent unit network to extract time dynamic features; S300-3, concatenating the feature vectors output by the two CNN branches with the feature vector output by the GRU network, and sending them to a multi-head self-attention mechanism layer for weighted fusion to adaptively emphasize key features; S300-4, based on the fused features, outputting the fault type, severity level and diagnostic confidence through a classifier.

4. The partial discharge fault identification method for a high-voltage switchgear according to claim 1, characterized in that, Step S400 includes the following steps: S400-1, constructing a digital twin model of the high-voltage switch cabinet, which integrates three-dimensional geometry, material physical properties and temperature field distribution; S400-2, setting a virtual discharge source candidate position in the digital twin model; S400-3, based on the physical properties of the digital twin model, simulating and calculating the simulation waveforms of the ultrasonic wave propagating from the candidate position to each sensor; S400-4, calculating the average coherence in the frequency domain between the measured ultrasonic wave signals and the simulation waveforms on each sensor channel; S400-5, taking the maximization of the average coherence as the objective function, and iteratively updating the virtual discharge source candidate position using a particle swarm optimization algorithm; S400-6, determining the optimal candidate position at the end of iteration as the actual position of the partial discharge source.

5. The partial discharge fault identification method for a high-voltage switchgear according to claim 1, characterized in that, Between step S300 and step S400, there is also step S600 of updating the cascade deep learning network based on a federated learning framework; wherein the multiple edge nodes train local models on local data and upload model weights to the cloud, and the cloud server generates a global model after securely aggregating the weights and distributes it to each edge node.

6. The partial discharge fault identification method for a high-voltage switchgear according to claim 1, characterized in that, In step S500, there is also a risk assessment step: S500-1: receiving the fault type, severity level and positioning information; S500-2: combining device historical working condition data, real-time load and environmental factors, calculating short-term and long-term risk probabilities of fault evolution into accidents through a Bayesian risk network; S500-3: based on the risk probability, automatically generating differentiated operation and maintenance strategies.

7. The partial discharge fault identification method for a high-voltage switchgear according to claim 4, characterized by, In step S400-1, the digital twin model receives real-time temperature data and updates the temperature field distribution inside the model, and then dynamically adjusts the ultrasonic wave propagation speed in the acoustic simulation.

8. The partial discharge fault identification method for a high-voltage switchgear according to claim 2, characterized in that, In step S200-1, the adaptive noise complete set empirical mode decomposition method is used to decompose the signal into multiple intrinsic mode functions, and the IMF components related to the partial discharge characteristics are selected for reconstruction to achieve signal denoising.

9. The partial discharge fault identification method for a high-voltage switchgear according to claim 1, characterized in that, In step S100, the acquisition process of the multi-physical field sensor array is controlled by a precise clock synchronization unit, ensuring that the synchronization accuracy of all signals is in the microsecond level.

10. A computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method of any one of claims 1 to 9.

Images